Image Registration

Question 1

What can the meta information of a image contain?

Answer 1

Scale, Orientation, Position…

Question 2

What is a rigid transformation and how many degree of freedoms does it have?

Answer 2

Rotation + translation, 3/6

Question 3

What is a similarity transformation and how many degree of freedoms does it have?

Answer 3

rigid + uniform scaling, 4/7

Question 4

What is a afine transformation and how many degree of freedoms does it have?

Answer 4

rigid + non-uniform scaling + shear, 6/12

Question 5

How can we perform a shear transformation?

Answer 5

HSH^-1 where H is rotation and S is scaling

Question 6

What uses has image registration?

Answer 6

Panorama stitching, optical flow, comparission (e.g. medical)…

Question 7

What is the assumption behind SSD (sum of squared differences) and SAD (sum of absolute differences) cost functions for intensity registration?

Answer 7

1. Identity relation between intesity distributions

2. Mono modal

Question 8

What is the assumption behind CC (Correlation coefficient) cost functions for intensity registration?

Answer 8

1. Linear relation between intesity distributions

2. Mono modal.

Question 9

How can we create a cost function using 2d histograms of intensities for intensity registration?

Answer 9

Search for a registration with where the 2d histogram has most high-value bins. The most used cost function is:
D_NMI = - NMI = (H(I) + H(J)) / H(I,J)

Question 10

What is the assumption behind D_NMI (Dissimilarity Normal Mutual Information) cost functions for intensity registration?

Answer 10

1. Statistical relation

2. Can be multi modal

Question 11

Where in two images is the disimilarity meassured

Answer 11

Only in the overlap of the registration

Question 12

How can we leviate the problem of local minimas in Image registration?

Answer 12

Do registration on pyramids of gaussians.

Question 13

What are some heuristics for image registration initialization?

Answer 13

1. Initialize to center of images

2. Inaitalize to center of mass

Question 14

What kind of search do we usually use for image registration?

Answer 14

Iterative search. (Gradient descent, bayesian optimization…)

Question 15

What kind of architectures do we often use for image registration with neural networks?

Answer 15

Convolutional similar to architectures used for image segmentation.

Question 16

What kind of datasets can be used for training of neural networks in image registration?

Answer 16

Syntetic data, (Computer generated video, Flying chairs…)

Question 17

What is the FlowNet 2.0?

Answer 17

A network for optical flow using a 3 network pipeline for large displacement and a parallel network for small displacement.

Question 18

What is the idea behind a spacial transformer network?

Answer 18

Add a network that learns parameters for a spacial transform before a classification network. The spacial transform is performed to the input of the classification network. The transform network is not trained directly, but trained by backpropagating classification errors.

Question 19

What is the idea behind the TeTrIs network?

Answer 19

Feed the image and a shape prior to a neural network that finds the transformation parameters. Perform the transformation on the prior.

Question 20

What is the formula for CC (Correlation coefficient)

Answer 20

covariance / (sigma_1 * sigma_2)